What is Draft Model?

Speculative decoding uses a fast draft model to generate candidate continuations which the larger target model then checks in parallel, accepting what matches and correcting what does not. The output is identical to what the large model would have produced alone; the gain is speed. It is a serving optimization and does not change behaviour, which makes it largely invisible to an organization consuming hosted inference. Its indirect relevance is the same as every efficiency improvement: cheaper inference encourages more of it, and more calls mean more opportunities for ungoverned requests. Centralpoint's controls operate independently of serving efficiency — the retrieval surface is bounded at index time, governance skills load ahead of retrieved content, and consumption is metered per execution, so a faster model produces faster governed answers rather than a larger ungoverned surface.