Correct. But the 5% more solves and 3-5% fewer tokens are within the noise of previous eval runs I've done on models. The faster response time is not.

Someone using an actual statistical test to quantify if there's actual measurable improvement?! An angel among men.

from my basic stats knowledge, it's not reliable to calculate the confidence interval of the sample proportion with fewer than about 10 samples on both sides. So if you're only seeing 1 or 2 failures in 20, and that approximately held as you scaled, you would have to run about 200 tests with each in order to get at least 10 failures in order to calculate the confidence intervals of the two samples and see if they overlap or not.

And this is why not many people do this. Even if you have the decals and the statistical knowledge, it's pretty expensive to do this regularly.

Uncertainty from count statistics are far larger than this. These are two perfectly consistent results there’s no statsig difference