Benchmarks

1. DTLZ5 / DTLZ6 apply the angular \(\theta\) map only to the first \(M-1\) decision variables. Distance variables feed \(g\) and are not multiplied into \(f_1\) as extra cosines.
2. Chuang F3 scores the selector-1 branch with the published trap (not the inverse trap) and covers bits \(0..39\) without dropping bit 38. The wrap block is concatenated, not added elementwise, so a NumPy individual scores the same as a list.
3. Moving Peaks ALT1 uses move_severity \(1.5\), matching its documented table. A pfunc pool shorter than npeaks fills every peak from the pool instead of raising ValueError, and a negative uniform_height or uniform_width draws random values, as documented for <= 0, instead of being used as is.
4. Royal Road R2 sums R1 at every doubling of order that still fits the bit string. The top-level schema is included, so a 64-bit all-ones individual with order \(8\) scores \(256\).
5. Kotanchek uses the published denominator $1.2`, not \(3.2\).
6. Royal Road R1 (and R2, which sums R1) and bin2float treat boolean bits like integers. mut_flip_bit preserves bool; bits no longer stringify as "True" / "False" and raise ValueError.
7. Scaled Rastrigin accepts a one-dimensional individual. The \(10^{(i-1)/(N-1)}\) scale is \(1\) at \(N = 1\) instead of raising ZeroDivisionError, so it matches plain Rastrigin.