One Tick, One Link: A Non-Autoregressive Challenge–Response Memory in Which Depth Is Bought with the World's Silence

We report a sequence learner that is not autoregressive. It is challenge–response: one input per tick, one output per tick, and intervals in which the input is the identity and the model keeps emitting. Its two operations are inverses selected by the world — speech binds an observation into a superposed store, silence unbinds one link out of it — so that depth is neither a layer count nor a hyperparameter but the number of quiet ticks the world grants. Several partial responses mature in parallel and the one that has matured is said. There is no reward, no label, and no gradient across a memory write; the only external quantity is prequential codelength. We verify in isolation that a chain is traversed one link per quiet tick and halts at its end, derive and confirm the store’s operating region d > 2k ln V , show that binding is invertible only for unitary codebooks, and show that locality-sensitive addressing makes a near miss undetectable by construction. On a synthetic continual stream retrieval occurs and holds, and parallel responses are load-bearing. We rule on none of the components that measure net-negative, because the explanation for all of them rests on a property of our source rather than of the world, and because a replayed file cannot test a system for which time is a resource.