Work out annotation qa sample size instantly with clear inputs, formula shown and shareable results.
Zero-acceptance sampling asks how many items you must inspect to be confident of catching at least one defect if the batch is at or above a given error rate. The probability of missing every defect in n draws is (1-p)^n, so n = ln(1 - confidence) / ln(1 - p). Detecting a 2 percent error rate with 95 percent confidence needs about 149 items, regardless of whether the batch is 5,000 or 500,000.
Zero-acceptance sampling
n = ln(1 - confidence) / ln(1 - p); detection probability = 1 - (1 - p)^n; reject the batch on any defect found
This plan is designed to reject on any defect, which is what gives the small sample size. Allowing defects requires a larger sample from a binomial or AQL table.
Yes, and drawn across annotators and time within the batch. Reviewing the first hundred items measures the warm-up period rather than the batch.