The eval loop is the beating heart of a quality improvement process.
For example, a search quality problem.
The loop: sample sessions that had a bad result, come up with scalable ways of improving them, experiment, ship, repeat.
At the very beginning it's hard to get that eval loop humming.
The eval loop can absorb all of the attention and resources you give it.
So be careful to only give it the proper amount of attention.
If the core product has PMF and you're still getting super-linear returns from the loop, then keep investing more in it.
But if the eval loop is for a secondary part of the product, or a product that doesn't yet have PMF, or is getting significantly diminishing returns because you're hitting the quality asymptote, pull back resources.