Rollout
Running a prediction forward more than one step by feeding each answer back in as the next question.
Like you're twelve
Before you play a move in chess you imagine it, then imagine the reply, then your answer to that. Each guess rests on the one before, which is why thinking six moves ahead is much harder than thinking one — a small mistake early makes everything after it wrong.
That chain of imagined steps is a rollout, and the reason your chess brain gets fuzzy at depth is the same reason a memory does.
Precisely
A rollout applies a world model repeatedly: predict the next state, then predict from that prediction. Each step's error becomes the next step's input, so accuracy decays with the horizon — a rollout is a compounding of approximations, not a lengthening of one.
A cleanup step between steps is what makes any depth reachable. Without it, each prediction carries the last one's noise forward and the chain degrades within a few steps. With it, each step is snapped back onto a stored state before continuing, which resets the noise floor rather than accumulating it.
How far a rollout survives is set by how crowded each read is rather than by depth itself. In a pool of 16 items, six chained steps recover perfectly; in a pool of 128, the first step is already unreliable — per-step recall multiplies, so capacity is the real limit.