Comment by buildbot

It can be exact or not! Depends on the kind of sampling you are doing.

You can do exact verification, and as soon as a token mismatches you reject everything after that token from your draft. Relaxed acceptance techniques measure how wrong that mispredicted token is via some metric, and accept it if it’s close enough. So you get longer draft lengths with higher acceptance rates.