We corrected ourselves in public
Our earliest headline benchmark claimed the AVX-512 kernels ran 9.15× faster than scalar code. A number that good deserved suspicion, so we rebuilt the measurement: 50 samples, dual runs, a frozen and documented protocol. The result came out at 1.27×. We published the correction ourselves, and both numbers stay visible on this page. That correction is the founding story of this project: evidence over flattery, including our own.