Models were always generating facts in order. The information was always there — streaming tokens proved that. What was missing was a protocol that exposed that sequence as structured facts and accepted injections at the boundaries between them. Patch streaming is that protocol. It does not add a capability to the model. It exposes a capability the model always had, by providing a yield surface where none existed before.
(An LLM with patch streaming is a coroutine · The fundamental unit of LLM output is a fact, not an object · JSON’s closing brace is an unexamined synchronization point)