Estimate a confidence interval for the true accuracy of a labeled dataset from a reviewed sample. Enter the number of labels audited, the number judged correct, and the confidence level. The calculator reports observed accuracy, a Wilson score interval, margin of error, and estimated error count—useful when deciding whether a batch is ready for training or needs more quality review.
The interval describes values that are reasonably compatible with the audited sample under binomial sampling. It does not correct for a biased audit sample or inconsistent reviewers.
This is a statistical estimate, not proof that every production label meets the same accuracy level. Sample randomly and audit independently.
With 500 labels reviewed and 460 correct, observed accuracy is 92.00%. At 95% confidence, the Wilson interval is about 89.29% to 94.07%, with an approximate half-width of 2.39 percentage points.
The Wilson interval behaves better than the simple normal interval when samples are small or accuracy is near 0% or 100%.
Yes, but the interval is most useful when the reviewed records are an independent, representative audit of a larger batch.
The upper bound reaches 100%, while the lower bound remains below 100% to reflect remaining sampling uncertainty.
Not in the strict frequentist sense. It means the method captures the true proportion in about 95% of repeated comparable samples.
Review more independent labels, keep the sampling representative, and avoid using a higher confidence level unless it is required.
| Input | Unit | Role |
|---|---|---|
| Labels reviewed | labels | The independently audited sample size. |
| Labels correct | labels | Labels that passed the audit. |
| Confidence level | % | Common choices are 90%, 95%, and 99%. |