Correlations. Period. For example, optical Bell tests generate a correlation between rate of coincidental detection and the angular difference between the analyzers. The observed correlations are in line with what would be expected given the validity of certain standard optical principles which are part of the standard qm treatment. The local realist correlations differ from this, and the point of departure is the requirement of formally explicating realism and locality. The now conventional way of doing this, a la Bell, is necessarily incompatible with the standard qm treatment and established optical principles which are part of that treatment. So Bell inequalities are violated, and it's neither mysterious nor deep why they're violated. Nevertheless, insofar as it can be considered the general LR form, then the LR program has been effectively refuted. But this doesn't inform us about the reality underlying the instrumental results. And, given that nonlocality hasn't been empirically or reasonably demonstrated, then, apparently, LR theories of entanglement are impossible even in an exclusively locally evolving universe.
The ONLY reason that the idea of nonlocality gained a foothold is because it provides an ad hoc answer to how the observed correlations might be accounted for in a realistic, formally separable model of entanglement.
One key point here is that the formal nonseparability of standard qm doesn't imply or refer to nonlocality. So, continuing to use the term nonlocality in this way just perpetuates the confusion.