Take only real examples from real review queues. Don't tell me there aren't enough real rejected edit suggestions and the system has to make up absurdly unbelievable ones.
Present the audit exactly as it appeared when it first entered the review queue, with the same text, comments, votes, closure reasons, etc.
Test the audit samples before use by presenting them to several more reviewers. Only samples that indicate a strong consensus should be used for audits.
Measure audit sample performance in use: if more than, say, 10% of reviewers get it "wrong", scrap it.