Improve evaluation pipeline SQL testing

[RamiAwar]

Privileged issue

• I'm @RamiAwar or he asked me directly to create an issue here.

Issue Content

Eval pipeline SQL tag doesn't look too stable across all runs. We should look into the cause of instability - why is it such a different score from one run to another?

If all else fails, consider aggregating statistics to reduce noise.