I am guessing here.

Corrections based on larger context should also be part of the streaming output -- maybe include replacement text for previous chunk/s identified by chunk ID.