The language model is the creative part of tailored and the only part with no authority over what ships. It reads the facts and the posting, then proposes prose, a summary, reworded bullets, a cover note in the candidate's voice. That is all it does, and it does it a little differently every run.
That variance is both the point and the danger. The same prompt drifts somewhere new each time, so the output is plausible every run and dependable on none of them. A sharper prompt does not fix this, because the next run drifts somewhere else instead. What fixes it is refusing to let the drift reach the page.
So the model sits inside a fence it cannot move. It proposes; the gates decide. Everything downstream of this node exists to catch what a fluent, ungrounded writer would otherwise wave through. It is the one un-anchored node in the graph, which is exactly why the anchored ones surround it. The longer argument for that shape is in Loops, all the way up.