An easy policy to choose and enforce is to simply disallow MI responses to yes-no questions. It may seem harsh and inflexible, but the solution has merit:

• Yes-no questions become simpler and more consistent
• Grammars become smaller and more robust at rejecting OOG
• Users learn the rules of the dialogue quickly

If this is your strategy, then the first and most important design feature is to ensure extremely high rejection of OOG. Build a yes-no grammar that has high performance on in-grammar "yes" and "no" answers. Then avoid adding MI phrases for coverage. Instead, treat all OOG as "not a yes or a no" and recover the error quickly by insisting on a yes-no answer.

Are you a member of this club?

"I'm not a member, but my father is, and he asked me to call."

(ASR returns OOG)

Please answer yes or no. Are you a member of this club?

In other words, since we're disallowing MI responses, we don't need to recognize them. Instead, we have to build a simple automaton that has very strict yes and no rules. We have to make sure that the automaton behaves properly when the user follows those rules. And finally, we have to ensure that the automaton teaches users quickly and firmly when they don't know or don't follow the rules.

Most callers are one-trial learners and they'll do just fine.