1 · SelectionFrom the timeline so far, the machine walks down its tree of choices by UCT, Q + c·√(ln N / n): the branches that have paid off, balanced against the ones it has barely tried.
2 · ExpansionAt the tree's edge it opens a new branch: one of the eight next pieces the archive offers here, found by CLIP embedding, the strict judge's score, and the story (the beat it plays, the pictures that answer the card before it).
3 · RolloutFrom that branch it imagines the rest of the film, quickly and a little by chance, all the way to The End. Each rollout is a small film.
4 · BackpropagationThe critic scores each imagined film (Andersonian frames, matched cuts, cards borne out, a story in order, pace, music) and the reward flows back up every branch that led to it. After 128 futures the most-visited piece is committed, and the search goes on from there.