...in that framework, how can an LLM - a statistical reassembler of things we already know about - evaluate a research output?

You know the answer already: it cannot.

Ipso (w00t science word!), an academia using these systems is lost. And must be abandoned - recast into something new.

/fin