@seldo The OSCE comparison is where I'd push. A standardized patient is trained to answer the same way for every student, and that's what makes the checklist fair. An LLM customer drifts: the same model can volunteer a hidden requirement to one candidate and hold it back from another who asked nearly the same thing. I'd log which requirements the customer revealed in each session, and score candidates on the questions they asked rather than on what the customer happened to give away.