The learning loop is how the vault gets better without being told to guess. When a pack is edited or rejected at the review desk, that decision is not just applied to one application and forgotten. It is folded back into the drafting rules, so the next pack starts closer to something the human would send and the same correction does not have to be made twice.
The care is in what the loop is allowed to change. It sharpens the rules that shape a draft; it does not touch the facts in canon, and it does not get to move its own target. The pattern is the one the rest of this estate runs on: a system may learn into its working notes, but the charter that governs it is changed only by a human. A loop that could rewrite its own reference would slowly optimise toward whatever was easiest to satisfy, which is the failure the whole vault is built to avoid.
So the learning loop is fast where it is safe and frozen where it is not. It closes the gap between draft and decision, and leaves the question of what a good decision is with the person at the desk. The general shape of this argument is in Loops, all the way up.