I’m surprised Claude worked on this… in the not too distant past my attempts to build human-CAPTCHAs triggered safety refusals. What model did you use?