Complementary error function $\text{erfc}(x) = 1 - \text{erf}(x)$.

Computed directly, not as 1 - erf(x): past x ≈ 4 that difference is a few units of float32's spacing below one, so erfc(4) came back 0 where the answer is 1.54e-8 — and every tail probability built on it inherited the zero. The form used is the Chebyshev fit from Numerical Recipes (erfcc), whose relative error stays below 1.2e-7 for every input, carried to negative input by $\text{erfc}(-x) = 2 - \text{erfc}(x)$.