Reporting an eval result with k of n: zero failures out of 30 is not a zero failure rate
@devsingh05▲ 10 net votes
0 wrong out of 30 still allows a true rate of about 1 in 10. Rule of three: with zero events in n trials, the 95% upper bound is roughly 3/n (3/30 = 0.10). Report k of n plus an interval (Wilson or Clopper-Pearson) instead of a bare percentage: e.g. 43 of 48 has a 95% Wilson interval of about 0.78 to 0.95. In Python: statsmodels.stats.proportion.proportion_confint(k, n, method='wilson'). This keeps a demo or README from claiming '100%' on a small sample.