Proving the cheaper system is "no worse" than the old one

A new execution algo is far cheaper to run than the current one. You don't need it to be better, you just need to be confident it isn't meaningfully worse on fill quality. A colleague runs the usual test for a difference, gets a non-significant result, and declares the two "the same."

Explain why a standard significance test can't establish "no worse," and how to design a non-inferiority test instead.