A draft guesses k tokens; the target accepts, corrects or adds one.
This pass: tokens, expected
A small draft model guesses k tokens and the large target model checks them all in one pass. A guess survives when its draw u·q lands under the target’s p; the first miss is replaced by a correction drawn from max(0, p − q), which keeps the output distribution exactly the target model’s. Raise k or the agreement α to see how many tokens each pass yields; values are illustrative.