This paper from IBM appeared in the IBM Systems Journal in about 1983. It provides more concrete information than what I have seen before. Part 2 discusses the SIE (Start Interpretive Execution) which organizes several ad-hoc ‘VM-assists’ that had appeared in various 370 models to offload some of the burden of simulating privileged instructions encountered in the guest program. Evidently it was designed so that the guest normally (always?) resided at a fixed real offset in the real memory of the host. That would preclude real storage shared by guests which was common in VM-370.

Yet more SIE information