Eval Suite Pass Rates

Eval suite pass rate vs suite size

A suite passes only when every case passes. Each case passes independently with probability p. With retries, a case counts as passing under pass@k if any of k attempts passes, and under pass^k only if all k attempts pass.

%
Plot

pass@k · retry until one passes

Case passes with 1 − (1 − p)k; suite passes with (1 − (1 − p)k)N

pass^k · every attempt must pass

Case passes with pk; suite passes with pkN

Assumes every case has the same pass rate and attempts are independent. Real flaky cases are correlated (the same prompt tends to fail the same way), so retries help less than pass@k suggests. The dashed vertical line marks your suite size; the table reads values at that size.