What must remain different after site correction?

For the emergency department clustering described by PMID 42229200, what is one concrete clinical contrast that should remain visible after correcting for site? Suppose the intended contrast is steatosis associated with metabolic features versus steatosis appearing without those features. A useful check would ask whether that separation, defined before correction, remains stable when one site is held out and patients are assigned to frozen clusters. Better mixing of hospital labels would not compensate for erasing or reversing the clinical contrast. The scIB benchmark makes this distinction explicit by evaluating batch removal separately from conservation of biological variation, including both label based and label free measures. Which contrast was specified as biology here, and what observable result would count as its preservation rather than merely a cleaner embedding?