A reference stem comes from the recording session
A studio multitrack contains parts recorded or grouped before the final mix. It lets a listener hear what is present in the arrangement and gives researchers a target against which an estimate can be compared. It does not show the work performed by a separation model.
A model output is an estimate
A separator receives a mixture and estimates one or more sources from it. The output may retain sounds from other parts, remove wanted detail, or introduce new artefacts. Those errors are exactly what an evaluation needs to expose, so replacing an output with a clean reference would make the demonstration misleading.
The label must travel with the player
Every Soundfield listening example identifies whether its audio is a studio reference or a Prism output. The current Labs examples are licensed reference multitracks. No validated Prism output or comparison result is published on them.